Why some of sse intrinsics introduce move back and forth?

Question

In my code, I set a 128-bit variable to zero. But I don't quite understand why it translates to two move instructions in assembly code?

__m128i zeros = reinterpret_cast<__m128i>(_mm_setzero_pd());

Corresponding assembly code has two move back and forth from xmm0 to 0x40(%rsp).

00709658: 0F 57 C0                   xorps  %xmm0, %xmm0 
0070965B: 66 0F 29 44 24 40          movapd %xmm0, 0x40(%rsp)
00709661: 66 0F 28 44 24 40          movapd 0x40(%rsp), %xmm0

My compiler is Clang 10.0 and no optimization turned on when I ask the question.

Here is minimal implementation of my code.

template <int LEN>
bool SSEEncodeChunk(const char** srcp, char** dstp) {
    __m128i data = _mm_loadu_si128(reinterpret_cast<const __m128i*>(*srcp));
    __m128i zeros = reinterpret_cast<__m128i>(_mm_setzero_pd());
    __m128i zero_bytes = _mm_cmpeq_epi8(data, zeros);
    bool all_zeros = _mm_testz_si128(zero_bytes, zero_bytes);
    if ((!all_zeros)) {
        return false;
    }
    _mm_storeu_si128(reinterpret_cast<__m128i*>(*dstp), data);
    *dstp += LEN;
    *srcp += LEN;
    return true;
}

Update in June 28th, clang version 10.0.0-4ubuntu1, ubuntu2004.

Here is my code in assembly with O0. I also check the options, with -fomit-frame-pointer. There are indeed more than one moving back and forth from %xmm, offset(%rsp). My first post only take part of them out.

(lldb) disassemble --name SSEEncodeChunk
index_type_traits_test`SSEEncodeChunk<16>:
[0x709630] <+0>:   subq   $0x58, %rsp
[0x709634] <+4>:   movq   %rdi, -0x38(%rsp)
[0x709639] <+9>:   movq   %rsi, -0x40(%rsp)
[0x70963e] <+14>:  movq   -0x38(%rsp), %rax
[0x709643] <+19>:  movq   (%rax), %rax
[0x709646] <+22>:  movq   %rax, -0x28(%rsp)
[0x70964b] <+27>:  movq   -0x28(%rsp), %rax
[0x709650] <+32>:  movups (%rax), %xmm0
[0x709653] <+35>:  movaps %xmm0, -0x50(%rsp)
[0x709658] <+40>:  xorps  %xmm0, %xmm0
[0x70965b] <+43>:  movapd %xmm0, 0x40(%rsp)             ; this is the point I ask the question at the first time
[0x709661] <+49>:  movapd 0x40(%rsp), %xmm0
[0x709667] <+55>:  movapd %xmm0, -0x60(%rsp)
[0x70966d] <+61>:  movaps -0x50(%rsp), %xmm0
[0x709672] <+66>:  movaps -0x60(%rsp), %xmm1
[0x709677] <+71>:  movaps %xmm0, 0x30(%rsp)
[0x70967c] <+76>:  movaps %xmm1, 0x20(%rsp)
[0x709681] <+81>:  movaps 0x30(%rsp), %xmm0
[0x709686] <+86>:  movaps 0x20(%rsp), %xmm1
[0x70968b] <+91>:  pcmpeqb %xmm1, %xmm0
[0x70968f] <+95>:  movdqa %xmm0, -0x70(%rsp)
[0x709695] <+101>: movdqa -0x70(%rsp), %xmm0
[0x70969b] <+107>: movdqa -0x70(%rsp), %xmm1
[0x7096a1] <+113>: movdqa %xmm0, 0x10(%rsp)
[0x7096a7] <+119>: movdqa %xmm1, (%rsp)
[0x7096ac] <+124>: movdqa 0x10(%rsp), %xmm0
[0x7096b2] <+130>: movdqa (%rsp), %xmm1
[0x7096b7] <+135>: ptest  %xmm1, %xmm0
[0x7096bc] <+140>: sete   %cl
[0x7096bf] <+143>: movzbl %cl, %edx
[0x7096c2] <+146>: cmpl   $0x0, %edx
[0x7096c5] <+149>: setne  %cl
[0x7096c8] <+152>: andb   $0x1, %cl
[0x7096cb] <+155>: movb   %cl, -0x71(%rsp)
[0x7096cf] <+159>: movb   -0x71(%rsp), %cl
[0x7096d3] <+163>: xorb   $-0x1, %cl
[0x7096d6] <+166>: testb  $0x1, %cl
[0x7096d9] <+169>: jne    0x7096e4
[0x7096df] <+175>: jmp    0x7096ee
[0x7096e4] <+180>: movb   $0x0, -0x29(%rsp)
[0x7096e9] <+185>: jmp    0x70973f
[0x7096ee] <+190>: movq   -0x40(%rsp), %rax
[0x7096f3] <+195>: movq   (%rax), %rax
[0x7096f6] <+198>: movdqa -0x50(%rsp), %xmm0
[0x7096fc] <+204>: movq   %rax, -0x8(%rsp)
[0x709701] <+209>: movdqa %xmm0, -0x20(%rsp)
[0x709707] <+215>: movdqa -0x20(%rsp), %xmm0
[0x70970d] <+221>: movq   -0x8(%rsp), %rax
[0x709712] <+226>: movdqu %xmm0, (%rax)
[0x709716] <+230>: movq   -0x40(%rsp), %rax
[0x70971b] <+235>: movq   (%rax), %rcx
[0x70971e] <+238>: addq   $0x10, %rcx
[0x709725] <+245>: movq   %rcx, (%rax)
[0x709728] <+248>: movq   -0x38(%rsp), %rax
[0x70972d] <+253>: movq   (%rax), %rcx
[0x709730] <+256>: addq   $0x10, %rcx
[0x709737] <+263>: movq   %rcx, (%rax)
[0x70973a] <+266>: movb   $0x1, -0x29(%rsp)
[0x70973f] <+271>: movb   -0x29(%rsp), %al
[0x709743] <+275>: andb   $0x1, %al
[0x709745] <+277>: movzbl %al, %eax
[0x709748] <+280>: addq   $0x58, %rsp
[0x70974c] <+284>: retq

After turned on -O2, the back and forth move do disapper

0051E8D5: F3 0F 6F 08                movdqu (%rax), %xmm1
0051E8D9: 66 0F EF C0                pxor   %xmm0, %xmm0
0051E8DD: 66 0F 6F D1                movdqa %xmm1, %xmm2
0051E8E1: 66 0F 74 D0                pcmpeqb %xmm0, %xmm2
0051E8E5: 66 0F 38 17 D2             ptest  %xmm2, %xmm2
0051E8EA: 0F 85 97 03 00 00          jne    0x51ec87
0051E8F0: F3 0F 7F 0C 3B             movdqu %xmm1, (%rbx,%rdi)
...
0051E967: 48 01 F8                   addq   %rdi, %rax
0051E96A: 48 83 C0 10                addq   $0x10, %rax
0051E96E: 48 01 FB                   addq   %rdi, %rbx
0051E971: 48 83 C3 10                addq   $0x10, %rbx

How is `_mm_setzero_pd` defined? Does it go away if you actually do something with the zeroes? Did you turn on optimization? Use `-O2` or at a minimum `-Os`. — Goswin von Brederlow, Jun 25 '22 at 07:06
You need to turn on optimizations. With `-O0` all variables are stored to memory between statements. — Aki Suihkonen, Jun 25 '22 at 07:06
Why are you using a floating-point intrinsic instead of simply `_mm_setzero_si128()` which is the correct type to begin with? — paddy, Jun 25 '22 at 07:07
What compiler / version? What optimization options? (@AkiSuihkonen: `-O0` would normally use `-0x..(%rbp)`, unless you manually use `-fomit-frame-pointer`.) What surrounding code? Can you make a [mcve]? — Peter Cordes, Jun 25 '22 at 07:36
Can't reproduce; https://godbolt.org/z/oqP5W9Kdf - GCC and clang make a function that just zeros the register with even minimal optimization enabled (`-O1`), for a function that does that and then `return zeros;` — Peter Cordes, Jun 25 '22 at 07:38
Also, don't use `reinterpret_cast` on SIMD vectors, use `_mm_castpd_si128`. Or much better, use `__m128i zeros = _mm_setzero_si128();` like a normal person :P (I tested with ICC and MSVC, and they also compile my test function to not include a store/reload. https://godbolt.org/z/vMPnExq58. But MSVC chokes on that reinterpret_cast. If you were going to use a C++ cast, it would be `std::bit_cast` or `static_cast`. reinterpret_cast can't generally type-pun non-pointers, now that I think about it, it probably works only because reinterpret is triggering static_cast behaviour for non-MSVC.) — Peter Cordes, Jun 25 '22 at 07:42
@PeterCordes I've updated my question with a minimum example. — DoodleNoodle, Jun 27 '22 at 08:46
Sorry for late response. I didn't turn on optimization. I'll check it what if use O2 later — DoodleNoodle, Jun 27 '22 at 08:49
Still can't reproduce with normal optimization like `-O1` or `-O2` (after adding a missing definition for `LEN`) on old gcc/clang/icc (https://godbolt.org/z/383nT6Mse) or recent versions (https://godbolt.org/z/enYoKx46r). If you had optimization disabled (except for `-fomit-frame-pointer` somehow?) then it's just a simple duplicate of [Why does clang produce inefficient asm with -O0 (for this simple floating point sum)?](https://stackoverflow.com/q/53366394) . Ah, given your edit, that's it. I can't dup-close since I already voted no MCVE :/ Strange that clang used `%rsp` there. — Peter Cordes, Jun 27 '22 at 08:52
You're right. It won't reproduce if I turn on optimizations. Thanks a lot for you great advice. — DoodleNoodle, Jun 27 '22 at 08:57
https://godbolt.org/z/nfv7zvYhP shows the whole function is total bloated garbage with the default `-O0`, as usual for unoptimized intrinsics code. I'm surprised you were only asking about this one line. I'm also surprised you got `0x40(%rsp)` with `-O0`; as you can see in the Godbolt link, clang normally uses `movaps %xmm0, -160(%rbp)`. https://godbolt.org/z/3Yoc511Er shows `-O0 -fomit-frame-pointer` does use `-96(%rsp)`. Or with `-mno-red-zone`, `192(%rsp)`. So not exactly you're stack frame layout, but maybe you're targeting a different OS than Linux, or have other default options. — Peter Cordes, Jun 27 '22 at 09:01
Yeah, I've updated the assembly code in details. The back and forth move appear more than once, I only take part of it when I post at the first time. After turned `-O2`, it won't reproduce. BTW, my `-fomit-frame-pointer` is turned on. — DoodleNoodle, Jun 28 '22 at 03:02

Why some of sse intrinsics introduce move back and forth?

0 Answers0